An Ensemble Model for Enhanced Confidential Data Identification in Manufacturing Companies
Open Access DepositedManufacturing companies have data in multiple locations and multiple types of data, some of which are confidential. To protect growing volumes of unstructured text-based confidential data, organizations must first know where it is stored. This praxis proposes an ensemble ML model to enhance confidential data identification. The model takes a confidential file as input and finds other files in a data repository which may contain extracts of data or information from the original confidential file. The ensemble model consists of an embedding process to transform the files into vectors, an encoder which creates a latent representation of all files, and a classifier which identifies files which may contain confidential data and require further review and protection. This praxis compares embedding files using Bidirectional Encoder Representations from Transformers (BERT) and Longformer. For development of the encoder, this praxis compares standard stacked autoencoders, contractive autoencoders (CAEs), and Laplacian autoencoders (LAEs). For the classifier, a radius neighbor classifier is compared to k-Nearest Neighbors (kNN). This praxis contribution includes an analysis showing the validation loss of the autoencoder in the ensemble model does not correlate with the F1 score for the overall model. A second contribution is an analysis of the effectiveness of BERT and Longformer text embeddings in identifying confidential information and the execution time of each embedding technique. Finally, this praxis provides an analysis of combining a content consolidation technique with efficient transformer to overcome BERT’s 512 token limitation.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.